Anon (no.delete@this.spam.com) on January 20, 2020 3:35 pm wrote:
> Simon Farnsworth (simon.delete@this.farnz.org.uk) on January 20, 2020 2:40 pm wrote:
> > As an example, I could see a GPU driver developer deciding that they need to commit as much main memory
> > as VRAM available for each application, so that they are guaranteed to be able to move all VRAM objects
> > to system RAM (and hence to page it out if needed) if they need the VRAM for another application.
> >
>
> And what happens if the system overcommits and there is not
> enough memory when the video driver moves VRAM to RAM?

Then, at least on Linux, a process gets a SIGBUS signal, and has to deal accordingly with the OOM situation. The default behaviour is simply to die, but there are other alternatives available if you can handle a lack of RAM sensibly.

Also, in the particular case I described, that's a very unlikely situation - if you're committing a full video card's VRAM worth of address space for every process that uses Direct3D, but most processes don't fill VRAM with active objects, then it'll rapidly exhaust commit while using very little of that space.